If you don’t understand which foundation model to use for which job, you risk wasting time and money on the wrong tool — or worse, building something from scratch that already exists. Foundation models are pre-built, powerful AI engines that handle specific tasks like generating text, analysing images, or creating new visuals, and knowing the difference between PaLM, Gemini, and Imagen on Google Cloud is essential for passing the Generative AI Leader exam and making smart business decisions.
Jump to a section
A simple way to picture Foundation Models on Google Cloud (PaLM, Gemini, Imagen)
A busy restaurant kitchen at 7pm on a Saturday night. The head chef has three speciality assistants, each with their own unique skill set and tool collection, and the secret ingredient is that you can use them without knowing how to cook.
The first assistant, PaLM, is the text specialist. It reads menus, writes descriptions, and can translate the latest reviews into Spanish in seconds. It processes words like a language scientist, understanding context and nuance. You hand it a raw recipe idea in English, and it hands back a polished description in French.
The second assistant, Gemini, is the multimodal master. It can see a photo of the finished dish, read the recipe card, listen to the chef’s verbal instructions, and even analyse a video of the plating technique. It combines sight, sound, and text. It doesn’t just read the list of ingredients; it looks at the picture of the final dish and tells you what’s missing.
The third assistant, Imagen, is the visual artist. You describe the cake you want — three tiers, blue frosting, edible flowers — and it generates a photorealistic image of that exact cake, from scratch. It creates new visuals based on your description, no photography skills needed.
All three assistants are pre-trained. They come to work already knowing thousands of recipes, languages, and image styles. You don’t need to teach them how a sentence works or what a tree looks like. You just give them a specific task. That’s the power of foundation models on Google Cloud: they are ready-to-use, specialised AI tools that you can access without building or training a single model yourself.
Foundation models are large, pre-trained artificial intelligence models that serve as a starting point for a wide range of tasks. Think of them as a super-smart, well-educated assistant who has already read millions of books, seen billions of images, and learned the patterns of human language and visual data. You don’t need to train them from scratch, which saves enormous time, computing power, and money.
On Google Cloud, three primary foundation models stand out: PaLM, Gemini, and Imagen. Each is designed for different types of input and output, known as modalities. A modality is simply the form of data — text, image, video, audio, or code.
PaLM stands for Pathways Language Model. It is a text-only model. This means it takes text as input and produces text as output. It excels at tasks like summarising long documents, translating languages, answering questions, generating creative writing, and understanding sentiment. If you give it a customer review, it can summarise the main complaints and even suggest a response. PaLM is built on Google’s Pathways architecture, which allows it to learn across many different tasks efficiently. It is particularly good at reasoning, code generation, and maths problems.
Gemini is Google’s most advanced and largest model family. It is multimodal, which means it can process and generate multiple types of data — text, images, audio, video, and code — all within one model. You can give Gemini a video of a person explaining a concept, and it can both transcribe the speech and describe what’s happening visually. It can take a picture of a hand-written note and turn it into a typed email. Gemini represents the next generation of AI, combining understanding across different senses into a single AI. It is designed to be more nuanced and context-aware than PaLM.
Imagen is a text-to-image model. You give it a text description, and it generates a new, original image that matches that description. If you write 'a cat wearing a top hat in a steampunk library', Imagen creates that picture. It is used for creating marketing visuals, concept art, product mockups, and more. Imagen uses a diffusion process, where it starts with random noise (static) and gradually transforms it into a coherent image that matches the text prompt. It generates high-resolution, photorealistic images.
Why do these models exist? Traditionally, businesses had to build their own AI models from scratch, which required huge datasets of labelled information (like millions of photos of dogs all labelled 'dog'), massive computing resources (specialised hardware running for weeks), and deep expertise in machine learning. Foundation models remove this barrier. Google Cloud hosts these models and makes them accessible via simple Application Programming Interfaces (APIs). An API is just a way for one piece of software to talk to another. Instead of building your own translation AI, you make a call to the PaLM API, send it your text in English, and get back French. That’s it.
The key benefits are:
Speed: You can go from idea to working product in hours, not months.
Cost: You pay for what you use, not for the cost of training a model from zero.
Quality: These models are trained on enormous, diverse datasets and represent state-of-the-art performance.
Flexibility: You can customise them with your own data via a process called fine-tuning, teaching the model your specific jargon, style, or product catalogue without starting over.
Foundation models replace the old approach of building single-purpose AI from scratch. Before, a bank wanting a chatbot for customer service would have to build a new AI from the ground up. Now, they call Google’s PaLM API, give it their customer service scripts for fine-tuning, and they have a chatbot running in days. A marketing agency wanting a new ad image would have to hire a designer and a photographer. Now, they describe the image to Imagen and get it in seconds. A video analysis company would need to build separate AI models for speech recognition, object detection, and text analysis. Now, they use Gemini, which does all three in one model.
In summary, foundation models are the pre-trained, ready-to-use AI building blocks on Google Cloud. PaLM handles text, Gemini handles everything (text, images, audio, video, code), and Imagen creates new images from text. They exist to democratise AI, making powerful capabilities available to any business or developer without requiring a PhD in machine learning.
Identify your use case and modality
Start by deciding what you need to do: generate text, create an image, analyse a video, or a combination. This determines which foundation model to use. For example, if you need to write a product description, your input and output are both text (use PaLM). If you need to create an image from a text description, your input is text and output is an image (use Imagen).
Access Google Cloud and Vertex AI
Go to the Google Cloud Console and enable the Vertex AI API. Vertex AI is the central platform where foundation models are hosted. You will use Vertex AI to browse available models, test them, and generate API keys. This step requires a Google Cloud account and a project set up with billing.
Select the correct foundation model
From Vertex AI, choose PaLM 2 for text-only tasks, Gemini (Pro or Ultra) for multimodal tasks involving images, audio, video, or code, or Imagen for text-to-image generation. Each model is listed with descriptions and pricing. This selection is critical—using the wrong model will not work or will yield poor results.
Test the model with prompts (prompt engineering)
In Vertex AI’s Model Garden, you can test the model by writing a prompt (the instruction you give the AI). For PaLM, you write a sentence like 'Summarise this article in three bullet points'. For Imagen, you write 'A modern office with plants and natural light'. Testing lets you refine the output without writing code. You can adjust the prompt phrasing to get better results.
Fine-tune (optional) or deploy the model
If the out-of-the-box model is not accurate enough for your specific data (e.g., your company’s jargon), you can fine-tune it. Upload your labelled data (e.g., past customer support chats) and run a fine-tuning job in Vertex AI. This adjusts the model’s parameters slightly to improve performance. Otherwise, you can deploy the base model directly via API and start sending requests from your application.
Integrate the model via API and monitor usage
Write code (or use no-code tools) to call the model’s API from your application. For example, your e-commerce site calls PaLM API each time a product is added to generate a description. Monitor the number of calls, response times, and cost using Google Cloud’s monitoring tools. Adjust your usage to stay within budget and ensure the model meets performance needs.
An IT professional at a mid-sized e-commerce company is tasked with improving the customer experience across several touchpoints. The company sells handmade furniture. The IT lead, Sam, doesn't have a team of AI researchers, but they have a Google Cloud account and a deadline. Here is exactly what Sam does with foundation models.
First, Sam needs to improve the product descriptions on the website. Currently, the descriptions are short and inconsistent because the products come from different artisans. Sam uses the PaLM API. They write a simple script that takes the artisan’s raw description — 'a wooden table, brown, 4 legs' — and sends it to PaLM with a prompt like 'Write a compelling, three-sentence product description for a handmade wooden table...'. PaLM returns a polished, engaging description that includes the wood type, the craftsmanship, and the style. This process runs automatically for every new product uploaded.
Next, Sam wants to add a visual search tool. A customer can upload a photo of a room and the website should suggest furniture that matches. Sam builds a feature using Gemini, which can analyse the uploaded image, identify the colours, furniture styles, and even the room dimensions from the photo. Then it searches the product database for items that visually match. The customer uploads a picture of a rustic living room and the website immediately shows compatible rustic coffee tables and bookshelves.
Sam also builds a customer support chatbot. Again, Sam uses PaLM because the chatbot only handles text. Sam provides the chatbot with the company’s FAQ and return policy documents. The chatbot can now answer questions like 'What is your return policy for custom tables?' in fluent, natural language, 24 hours a day.
For a new marketing campaign, Sam needs visuals for a 'modern farmhouse' collection. Hiring a photographer and stylist for a full photoshoot would cost thousands and take weeks. Instead, Sam uses Imagen. They write prompts like 'A modern farmhouse dining room with a large oak table, white chairs, and a vase of sunflowers, photorealistic style'. Imagen generates several unique, high-quality images within minutes. Sam selects the best ones, makes minor edits, and uses them in the email newsletter and social media ads.
Finally, Sam wants to create a video tutorial on assembling a complex bookshelf. Sam records a screen capture and a voiceover, but the voiceover has background noise and the video needs visual captions. Sam uses Gemini to process the video. Gemini transcribes the speech into text, generates the captions, and can even suggest where to place text overlays. Sam also uses Gemini to translate the captions into Spanish and French to reach a broader audience.
The steps Sam took:
Identified the task: product descriptions, visual search, chatbot, ad imagery, video captions.
Matched each task to the correct model: PaLM for text tasks, Imagen for image generation, Gemini for video and mixed media.
Accessed the models via Google Cloud APIs, writing simple scripts or using no-code tools like Vertex AI.
Fine-tuned PaLM on the company’s product catalogue to improve accuracy.
Deployed the solutions and monitored usage, paying only for the API calls made.
Sam’s company saved months of development time, thousands of dollars in creative and development costs, and improved customer satisfaction scores. This is the practical, everyday reality of using foundation models on Google Cloud.
The Generative AI Leader exam tests your ability to identify which foundation model on Google Cloud is best suited for a given business use case. You will not be asked to write code or explain the transformer architecture. The exam is about decision-making, not deep technical implementation. Here is exactly what you need to know.
The exam loves scenario-based multiple-choice questions. You will read a short business scenario — for example, 'A marketing team needs to generate 50 unique social media images from text descriptions quickly' — and you must choose the correct Google Cloud foundation model from a list. The trap is that they will include similar-sounding models or vague options. The correct answer is almost always the one that matches the modality (text-only, multimodal, or text-to-image).
Key concepts the exam tests:
Modality: You must know what each model inputs and outputs. PaLM is text-in, text-out. Imagen is text-in, image-out. Gemini is any-in (text, image, audio, video), any-out (text, image, code).
Use cases: PaLM for chatbots, translation, summarisation, code generation. Imagen for creating marketing images, product mockups, concept art. Gemini for video analysis, mixed-media tasks, complex reasoning across data types.
Fine-tuning vs. prompt engineering: The exam will ask when to fine-tune a model (customising it with your own data for better accuracy on a specific domain) versus when to just write a better prompt (no model changes needed).
Vertex AI: The exam mentions Vertex AI as the platform to access, deploy, and manage these models. Know that Vertex AI is the central hub for foundation models on Google Cloud.
Cost implications: Larger models (Gemini Ultra) cost more per API call than smaller ones (PaLM 2). The exam may test your ability to recommend the most cost-effective model for a given task.
Common exam traps:
Offering 'GPT-4' or 'Claude' as options: These are not Google Cloud models. The exam only tests Google Cloud products.
Suggesting you build a custom model from scratch: Foundation models exist precisely to avoid this. The correct answer will always be to use an existing foundation model.
Confusing Imagen with a photo editing tool: Imagen generates new images from scratch, it does not edit existing photos. Google has a separate tool for editing (like AutoML Vision) but Imagen is creation-only.
Assuming Gemini replaces everything: While Gemini is multimodal, it is not always the best choice. For simple text-only tasks, PaLM is faster and cheaper.
Specific topics to memorise:
The three main models: PaLM 2 (text), Gemini (multimodal), Imagen (text-to-image).
That Gemini comes in three sizes: Ultra (largest, most capable), Pro (balanced), Nano (on-device).
That foundation models are pre-trained on large public datasets and can be fine-tuned with your own data via Vertex AI.
The difference between a 'base model' (out-of-the-box) and a 'tuned model' (customised with your data).
Ethical use: The exam may ask about responsible AI principles like avoiding bias in models, which is part of Google’s AI Principles.
Question pattern example: 'A news organisation wants to automatically generate audio descriptions of images for visually impaired readers. Which Google Cloud foundation model is best suited for this task?' The correct answer is Gemini, because it can take the image as input and generate text (description) as output. PaLM cannot process images. Imagen only generates images, not text. The trap answer would be 'Imagen' because a beginner thinks 'images, so Imagen', but the output is text, not an image.
Study the official Google Cloud documentation on foundation models and Vertex AI. Practise distinguishing between modality types. The exam is passable if you understand the 'what' and 'why' of each model, not the 'how' of building them.
PaLM is a text-only foundation model on Google Cloud, ideal for summarisation, translation, chatbots, and code generation, processing only text as input and output.
Gemini is a multimodal foundation model that can process and generate text, images, audio, video, and code simultaneously, making it the most versatile model for complex tasks.
Imagen is a text-to-image foundation model that generates brand-new, photorealistic images from text descriptions, used for marketing visuals and concept art.
Foundation models are pre-trained by Google on massive datasets, so you never need to train them from scratch; you can use them immediately via APIs on Vertex AI.
Fine-tuning allows you to customise a foundation model with your own data (like product catalogues) for better accuracy on your specific domain without retraining the whole model.
Choosing the wrong model for a task wastes time and money; always match the model's modality (text-only, multimodal, text-to-image) to your input and output requirements.
These come up on the exam all the time. Here's how to tell them apart.
PaLM 2
Text-only input and output
Best for tasks like summarisation, translation, chatbots
Lower cost per API call compared to Gemini
Gemini
Multimodal input (text, image, audio, video, code)
Best for complex tasks needing multiple data types
Higher cost per API call due to larger model size
Imagen
Generates new images from text prompts
Output is always an image
Used for creating marketing visuals and concept art
Gemini (image understanding)
Analyses and describes existing images
Output is text (description) or code
Used for image captioning and visual question answering
Fine-tuning
Requires uploading your own labelled dataset
Changes the model’s parameters permanently
Best for specialised domains with unique jargon
Prompt engineering
Only changes the input text instruction
Does not modify the model
Best for general tasks where the base model is already strong
Vertex AI Model Garden
Provides pre-trained foundation models ready to use
No machine learning expertise required
Pay only for API calls, no infrastructure management
Custom model training
Build a model from scratch using your own data
Requires deep ML expertise and significant compute
Higher upfront cost and ongoing maintenance
Mistake
PaLM can handle images because it's a large language model.
Correct
PaLM is a text-only model. It processes and generates only text. For images, you need Imagen (text-to-image) or Gemini (multimodal).
People confuse 'large language model' with 'can do everything'. Language models are text-focused. Multimodal models like Gemini are designed for multiple data types.
Mistake
Gemini is just a newer name for PaLM.
Correct
Gemini is a separate, more advanced model family that is multimodal, while PaLM remains a text-only model. They coexist on Google Cloud.
Google rebranded some products, so people assume a name change. But Gemini is a fundamentally different architecture from PaLM, not a rename.
Mistake
You need to train a foundation model from scratch with your own data before you can use it.
Correct
Foundation models come pre-trained on massive datasets. You can use them out-of-the-box or fine-tune them with your data, but you never train from scratch.
The term 'train' is misleading. Beginners think they must provide data to the model. Foundation models were trained by Google; you just use or adapt them.
Mistake
Imagen is used to edit or enhance existing photos.
Correct
Imagen generates brand-new images from text descriptions. It is not a photo editor. Google has different tools for editing existing images.
People see 'text-to-image' and think 'image editing' because many tools (like Photoshop) do both. Imagen is purely generative from scratch.
Mistake
All foundation models on Google Cloud are equally good at every task.
Correct
Each model is specialised. PaLM is best for text, Imagen for images, and Gemini for mixed media. Using the wrong model wastes money and produces poor results.
The marketing hype around AI suggests one model can rule them all. But in practice, choosing the right tool for the job is critical, just like a screwdriver is not a hammer.
Mistake
Using a foundation model is more expensive than building your own.
Correct
Building your own model from scratch is far more expensive due to compute costs, data collection, and expertise. Foundation models are pay-as-you-go and typically cheaper.
People see the per-call cost of an API and think 'permanent cost', forgetting the enormous one-time cost of training a model from zero.
Reveal each answer, then mark whether you got it right. Score 60%+ to unlock the next chapter.
No. Foundation models are designed to be used via simple API calls or no-code tools in Vertex AI. You just send a text prompt and get a response. No machine learning expertise is required to use them, though understanding the concepts helps you choose the right model.
Yes, Gemini can handle text-only tasks, but it is more expensive per API call than PaLM. For simple text tasks like summarisation, PaLM is the recommended and more cost-effective choice. Use Gemini only when you need multimodal capabilities.
Prompt engineering means changing the text instruction you give the model to get better output, without modifying the model itself. Fine-tuning means adding your own custom data to the model, actually updating its parameters to improve performance on a specific topic or style.
No. Imagen generates new images from text prompts. It does not edit existing photos. For editing, you would use other Google Cloud tools like Vision API or third-party software. Think of Imagen as an AI artist who paints from your description, not a photo retoucher.
Pricing is typically per API call, with some models having a base cost and a per-character or per-image cost. Gemini Ultra costs more than Gemini Pro, which costs more than PaLM. Imagen charges per generated image. You pay only for what you use, with no upfront costs.
No, the foundation models discussed (PaLM, Gemini, Imagen) are proprietary to Google Cloud and are only accessible through Google’s APIs. You cannot download them. You access them via the cloud, which means you need an internet connection and a Google Cloud account.
You've finished Foundation Models on Google Cloud (PaLM, Gemini, Imagen). Continue through the Generative AI Leader study guide to build a complete picture of the exam.
Done with this chapter?